8-1 Embedded Systems Fixed-Point Math and Other Optimizations

Embedded Systems8-2 Fixed Point Math – Why and How Floating point is too slow and integers truncate the data –Floating point subroutines: slower than native, overhead of passing arguments, calling subroutines… simple fixed point routines can be in-lined Basic Idea: put the radix point where covers the range of numbers you need to represent I.F Terminology –I = number of integer bits –F= number of fraction bits Bit Pattern Integer 0/1 = 0 28/1 = 28 99/1 = /4 = 0 28/4 = 7 99/4 = /1024 = 0 28/1024 = … 99/1024 = … 000 Radix Point Locations Bit Weight Weight21½¼1/81/161/32 Radix Point in a fixed point binary representation

Embedded Systems8-3 Rules for Fixed Point Math Addition, Subtraction –Radix point stays where it started –…so we can treat fixed point numbers like integers Multiplication –Radix point moves left by F digits –… so we need to normalize result afterwards, shifting the result right by F digits + * 10 * Format *

Embedded Systems8-4 Division –Quotient is integer, may want to convert back to fixed point by shifting –Radix point doesn’t move in remainder Quotient Remainder ÷ 7 ÷ Format ÷ Format ÷ Dividend Divisor

Embedded Systems8-5 Division, Part II Division –To make quotient have same format as dividend and divisor, multiply dividend by 2 F (shift left by F bits) –Quotient is in fixed point format now Quotient ÷ 7 ÷ Format (7*2) ÷ Format (7*4) ÷ Dividend Divisor

Embedded Systems8-6 Example Code for 12.4 Fixed Point Math Representation –typedef unsigned int FX_12_4; Converting to and from fixed point representation –#define FL_TO_FX(a) (unsigned int)(a*16.0) –#define INT_TO_FX(a) (a<<4) –#define FX_TO_FL(a) (float)(a/16.0) –#define FX_TO_INT(a) (unsigned int)(a>>4) Math –#define FX_ADD(a,b) (a+b) –#define FX_SUB(a,b) (a-b) –#define FX_MUL(a,b) ((a*b)>>4) –#define FX_DIV(a,b) ((a/b)<<4) –#define FX_REM(a,b) ((a%b))

Embedded Systems8-7 More Fixed Point Math Examples * * Format * * Format

Embedded Systems8-8 Static Revisited Static variable –A local variable which retains its value between function invocations –Visible only within its module, so compiler/linker can allocate space more wisely (recall limited pointer offsets) Static function –Visible only within its module, so compiler knows who is calling the functions, –Compiler/linker can locate function to optimize calls (short call) –Compiler can also inline the function to reduce run-time and often code size

Embedded Systems8-9 Volatile and Const Volatile variable –Value can be changed outside normal program flow ISR, variable is actually a hardware register –Compiler reloads the variable from memory each time it is used Const variable –const does not mean constant, but rather read-only –consts are implemented as real variables (taking space in RAM) or in ROM, requiring loading operations (often requiring pointer manipulation) A #define value can be converted into an immediate operand, which is much faster So avoid them Const function parameters –Allow compiler to optimize, as it knows a variable passed as a parameter hasn’t been changed by the function

Embedded Systems8-10 More Const Volatile Variables –Yes, it’s possible Volatile: A memory location that can change unexpectedly Const: it is only read by the program –Example: hardware status register

Embedded Systems8-11 Starting Points for Efficient Code Write correct code first, optimize second. Use a top-down approach. Know your microprocessors’ architecture, compiler (features and also object model used), and programming language. Leave assembly language for unported designs, interrupt service routines, and frequently used functions.

Embedded Systems8-12 Floating Point Data Type Specifications Use the smallest adequate data type, or else… –Conversions without an FPU are very slow –Extra space is used –C standard allows compiler or preprocessor to convert automatically, slowing down code more Single-precision (SP) vs. double-precision (DP) –ANSI/IEEE , Standard for Binary Floating Point Arithmetic Single precision: 32 bits –1 sign bit, 8 exponent bits, 23 fraction bits Double precision –1 sign bit, 11 exponent bits, 52 fraction bits –Single-precision is likely all that is needed –Use single-precision floating point specifier “f”

Embedded Systems8-13 Floating-Point Specifier Example No “f” float res = val / 10.0; Assembler: move.l -4(a6),-(sp) jsr __ftod //SP to DP clr.l -(sp) move.l # ,- (sp) jsr __ddiv //DP divide jsr __dtof //DP to SP move.l (s)+,-8(a6) “f” float res = val / 10.0 f; Assembler: move.l -4(a6),-(sp) move.l # ,- (sp) jsr __fdiv move.l (s)+,-8(a6)

Embedded Systems8-14 Automatic Promotions Standard math routines usually accept double-precision inputs and return double-precision outputs. Again it is likely only single-precision is needed. –Cast to single-precision if accuracy and overflow conditions are satisfied

Embedded Systems8-15 Automatic Promotions Example Automatic float res = (17.0f*sqrt(val)) / 10.0f; // load val // convert to DP // _sqrt() on DP returns DP // load DP of 17.0 // DP multiply // load DP of 10.0 // DP divide // convert DP to SP // save result two DP loads, DP multiply, DP divide, DP to SP conversion Casting to avoid promotion float res = (17.0f*(float)(sqrt(val)))/10.0f; // load val // convert to DP // _sqrt() on DP returns DP // convert result to SP // load SP of 17.0 // SP multiply // load SP of 10.0 // SP divide // save result two SP loads, SP multiply, SP divide

Embedded Systems8-16 Rewriting and Rearranging Expressions Divides take much longer to execute than multiplies. Branches taken are usually faster than those not taken. Repeated evaluation of same expression is a waste of time.

Embedded Systems8-17 Examples float res = ( 17.0f*(float)(sqrt(val)))/10.0f; is better written as: float res = 1.7f*(float)(sqrt(val)); D = A / B / C; as: D = A / (B*C); D = A / (B * C); as: bc = B * C ; E = 1/(1+(B*C)); D = A / bc; E = 1/(1+bc);

Embedded Systems8-18 Algebraic Simplifications and the Laws of Exponents Original ExpressionOptimized Expression a 2 – 3a + 2 2*, 1-, 1+ (a – 1) * (a – 2) 1*, 2- (a - 1)*(a + 1) 1*, 1-, 1+ a *, 1- 1/(1+a/b) 2/, 1+ b/(b+a) 1/, 1+ a m * a n 2 ^, 1* a m+n 1^, 1+ (a m ) n 2 ^ a m*n 1^, 1*

Embedded Systems8-19 Literal Definitions #defines are prone to error float c, r; r = c/TWO_PI; #define TWO_PI 2 * 3.14 –r = c/2*3.14 is wrong! Evaluates to r = (c/2) * 3.14 #define TWO_PI (2*3.14) –Avoids a divide error –However, 2*3.14 is loaded as DP leading to extra operations convert c to DP DP divide convert quotient to SP #define TWO_PI (float)(2.0*3.14) –Avoids extra operations

Embedded Systems8-20 The Standard Math Library Revisited Double precision is likely expected by the standard math library. Look for workarounds: –abs() float output=fabs(input); could be written as: if(input<0) output=-input; else output=input ;

Embedded Systems8-21 Functions: Parameters and Variables Consider a function which uses a global variable and calls other functions –Compiler needs to save the variable before the call, in case it is used by the called function, reducing performance –By loading the global into a local variable, it may be promoted into a register by the register allocator, resulting in better performance Can only do this if the called functions don’t use the global Taking the address of a local variable makes it more difficult or impossible to promote it into a register

Embedded Systems8-22 More about Functions Prototype your functions, or else the compiler may promote all arguments to ints or doubles! Group function calls together, since they force global variables to be written back to memory Make local functions static, keep in same module (source file) –Allows more aggressive function inlining –Only functions in same module can be inlined